For finetuning MBart-based model, setting decoder_start_token_id in model.config is NOT ENOUGH. #41492
Description
System Info
Context: finetuning a MBart model with run_translation.py
Easy fix is to set it in model.generation_config as well. Both worked outside of run_translation.py, but not setting this in run_translation.py causes validation/evaluation to fail miserably.
Who can help?
No response
Information
- The official example scripts
- My own modified scripts
Tasks
- An officially supported task in the
examplesfolder (such as GLUE/SQuAD, ...) - My own task or dataset (give details below)
Reproduction
Run run_translation.py to finetune with a validation file.
Expected behavior
BLEU scores should be terrible due to the bug. This makes it hard to monitor finetuning.